Anthropic accuses DeepSeek of Claude distillation

This is bigger than one lab calling out another — it frames frontier-model APIs as attack surfaces, not just developer products.

• –Anthropic’s claim matters because it says DeepSeek targeted the exact premium behaviors developers pay for: reasoning, tool use, and coding quality
• –The demo is useful because it makes “distillation attack” concrete: repeated high-quality outputs can become training data fast enough to narrow capability gaps
• –If this framing sticks, labs will tighten identity checks, behavioral monitoring, rate controls, and protections around reasoning traces
• –The article also ties distillation directly to export-control and national-security debates, which raises the odds of policy spillover into ordinary API access rules